At GTC this morning Nvidia finally pulled the wraps off its much-anticipated Ada Lovelace architecture along with three GPUs based on it. Jensen unveiled the GeForce RTX 4090 GPU and two RTX 4080s based on different dies. This is a first for the company and something that’s sure to confuse gamers. One is allegedly based on the AD104 die, with the other using the even-smaller AD103 die. The headline feature is the difference in memory — 16GB versus 12GB — but they have vastly different specs.
Jensen didn’t discuss the Ada cards’ details, but the website is already live. He noted Ada is up to 4X faster than the previous flagship in ray tracing for certain titles. The AD102 die will also have 18,000 CUDA cores, which is 70 percent more cores than the Ampere flagship. And just as the final teaser noted, the AD102 die has 75 billion transistors, a huge leap from Ampere’s 26 billion. The RTX 4090 is a triple-slot card, just like the RTX 3090 Ti. It also has the same TBP of 450W and looks almost exactly the same size. This lends credence to the notion the 3090 Ti was a test run of the upcoming cooling apparatus.
The RTX 4080 GPUs are much different than the RTX 4090. The 16GB version will be priced at $1,199, and the 12GB version will cost $899. They both offer lower core counts and narrower memory buses. The 16GB version is getting a 256-bit memory bus, compared with the 384-bit bus on the 4090. The 12GB version will be handicapped further with a 192-bit memory bus. This would give it a narrower bus than the RTX 3070, but it seems like Nvidia wants more separation between its GPUs this generation. The 16GB version will offer 9,728 CUDA cores, with the 12GB version getting 7,680. So far there is no mention of memory speeds for any of the GPUs, which is a bizarre omission.
As far as power consumption goes, the two RTX 4080s are pegged at 320W and 285W, respectively. Both the RTX 4090 and 16GB 4080 are three-slot GPUs that require three eight-pin PCIe cables, or a single 12+4 PCIe cable. The 12GB 4080’s size “varies by manufacturer” and requires just two eight-pin connectors or a single PCIe Gen 5 cable.
Ada seems like a pretty huge leap over Ampere, just as the rumors indicated. The RTX 4090 should be up to 4X faster in ray tracing than the RTX 3090 Ti, and the 4080 offers the same level of uplift over the 3080 Ti. We’ll have to see what the rasterization performance uplift is for Ada, but it sounds promising. The RTX boost seems to be coming from 3rd generation DLSS, or deep learning supersampling. Nvidia says it innovated a new rendering technique to use AI to predict frames for DLSS 3, helping to boost performance in games that are CPU or GPU-limited. Over 35 games will have support for DLSS 3, and it will be supported by Unity and the Unreal engine.
One notable highlight is Jensen said Ada has over 18,000 CUDA cores. However, the RTX 4090 has just 16,384. This indicates it’s got an RTX 4090 Ti waiting in the wings. We imagine that will appear if necessary to counter whatever AMD comes up with for RDNA3, which is launching on Nov. 3.
PNY’s VERTO RTX 4090 GPU looks swanky. (Image: PNY)
Possibly the biggest surprise is the lack of price gouging for Ada. Rumors had hinted at a price of over $2,000 for the RTX 4090, but the $1,599 pricing is surprisingly affordable given its theoretical performance. The same goes for the RTX 4080s, but we’ll have to wait and see how they perform against the existing 30-series. Power consumption is also surprisingly reasonable, but we don’t know if the cards will run hot. Nvidia’s website says their max temperature is 90C, but hopefully they won’t approach those toasty temps.
The RTX 4090 will be available on Oct. 12, with the RTX 4080s coming “in November.”
Now Read: